Notes on technology.
Partner at zaka.vc, investing early in health and industrial technologies. Having seen and read a great deal, and understood rather less of it, I am using this page to organise what stayed.
Writing
·zaka.vc·drug development
The US-China split now runs through ownership, not molecules
To compete with Chinese biotechs, you have to run as fast as they do, which takes a bigger round and more risk, or build something Beijing won't let them sell. China wins in drugs on time rather than labour cost: teams run toxicology, manufacturing and trial preparation at once, and do early discovery for a third to half of what it costs elsewhere. Beijing's draft rules from 16 October keep Chinese platforms licensable but take away the sale, and Washington's COINS draft limits American equity stakes. Pharma can rent a Chinese platform, but it cannot own one, so the question for a Western company is whether pharma will want to own it.
Read →·zaka.vc·personalized medicine
In a drug made for one patient, the AI is the product
AI in drug development mostly finds better candidates, which barely changes the cost of a drug, because most of that cost goes on trials. In medicines made for one patient, the trials, the approval and the manufacturing are shared across a platform, while the design has to be redone for each patient. That design work is mostly prediction, which AI suits, so the software sets the cost of each patient and the margin. The FDA's draft rules from February 2026 make such a business possible for the first time. The bet is that a platform gets cheaper with every patient it learns from.
Read →·zaka.vc·drug manufacturing
mRNA is still made the 1984 way. It was not built for the drugs coming next.
mRNA stayed a vaccine technology because a vaccine is a fifth of a milligram, given twice. A chronic protein-replacement therapy needs 0.4 to 1.6 grams a year, and a gram made to clinical standard costs between about $300,000 and $450,000. A bigger factory barely moves that, because RNA manufacturing runs on variable costs rather than fixed ones. Pharma has committed as much as $6 billion to RNA medicines that have to be dosed repeatedly, and none of the three announcements mentioned the price of a gram.
Read →·zaka.vc·clinical trials
You choose the science. The disease chooses the bill.
Most biotech founders never choose their first disease. It comes with the tissue their lab could get, and when asked, founders point to an early signal and a big market. Neither tells them what it costs to prove the drug works. Pivotal trials on an accepted surrogate averaged $24 million, and trials that had to show a clinical outcome averaged $65 million. The gap opens at the Series A, where few firms lead rounds big enough to close it.
Read →·zaka.vc·clinical trials
We now design more drugs than we can test
Discovery got much better at producing candidates and the step that tests them did not. Nearly four in ten adjuvant phase III oncology trials finish with less statistical power than their own design called for, and two thirds of trial starts now come from sponsors running their first or second study. The scarce input stopped being the molecule. It is a credible answer about one, and the CRO selling you the trial has no reason to hand you that.
Read →·zaka.vc·drug discovery
Nobody at a frontier lab is working on your disease
Frontier labs walking into drug discovery looks like an extinction event for seed-stage biotech. Alphabet ran the same experiment for twelve years across Calico, Verily and Isomorphic, and came up empty. Capital, patience and a Nobel-winning model were never the scarce inputs. Proprietary data and a program someone sees through still are.
Read →·zaka.vc·drug discovery
The moat decides whether you can play. The asset decides whether you get paid.
Discovery is priced before anyone knows the molecule works. Schrödinger co-discovered the compound Nimbus sold to Takeda for $4 billion and took home $147 million. A data moat tells you nobody can copy you. It says nothing about who pays you, and in target discovery the gap between those two is the whole business.
Read →·zaka.vc·drug discovery
In AI drug discovery, the model was never the moat
The question is no longer whether AI can design drugs but who keeps an edge once the frontier labs commoditize the model layer. The one test that survives: does a company manufacture biological data nobody else can. Archetype and business-model labels tell you little; data ownership tells you everything.
Read →